Google DeepMind and Singapore unveil scam detection model

Google DeepMind and Singapore's Government Technology Agency (GovTech) unveiled on October 14 a scam detection model that runs inside the smartphone. Built by fine-tuning Google's lightweight open model Gemma on local Singaporean scam conversations, the model reads messaging conversations on the device without sending them to the cloud and, when it sees manipulative intent, raises a real-time warning before money is transferred or sensitive information is shared. DeepMind said the model was optimized to run on the typical smartphones used in Singapore. The announcement came at Singapore International Cyber Week (SICW 2026).

The starting point is the scale of the harm. According to DeepMind's announcement, scams account for more than half of the crimes reported in Singapore, and cumulative losses since 2019 have exceeded S$4 billion. The tactics have changed too. Instead of simple phishing, social engineering scams that groom victims with convincing conversations before asking for money have grown, and such scams now account for up to 80% of all losses.

AI makes those conversations more convincing. Speaking the same day at the opening of the second day of Cyber Week, Josephine Teo, Singapore's Minister for Digital Development and Information, cited a case from earlier this year in which a victim was invited to a video meeting that appeared to include Prime Minister Lawrence Wong and senior officials. The faces were deepfakes, and the victim, a seasoned professional who believed he was contributing to a worthy global cause, sent nearly S$5 million. "Scammers have always exploited human trust. What AI changes is the cost of deception," Teo said.

The reason for putting the model on the device is encryption. Threat detection has relied on servers scanning messages in the cloud, but on end-to-end encrypted messengers even the operators cannot look into private conversations. DeepMind explained that moving threat analysis to the handset lets conversation patterns be analyzed faster, so users can stop before money or account credentials are exposed. According to reports, the model analyzes text conversations in messaging apps such as WhatsApp and Telegram in real time and is being developed for Android users. "Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, Director of Trust and Safety at GovTech, adding that it "allows us to catch highly targeted, local scam attempts in real time while keeping personal communications completely private."

Generic models struggle with deception techniques that vary by region. So Gemma was retrained on local scam conversations, ranging from impersonations of public institutions, banks and local figures to schemes trending in the region. The goal is to tell harmless chatter from genuine manipulation. METAL has reported that Google released an offline voice translator that runs on the device with Gemma, and this model puts the same lightweight open model to work as a security alarm.

Battery is the next hurdle. Running a complex classifier constantly on a phone would drain its power. DeepMind split the processing into two stages. In the first, a lightweight classifier that uses almost no power screens incoming messages and picks out only potential threats. In the second, the fine-tuned Gemma reads a larger part of the conversation and weighs the context to confirm a threat or discard a false alarm.

Where legal questions most easily arise is the training data. DeepMind set out three governance principles. The first is a sovereign sandbox: fine-tuning took place only inside a Google Cloud Platform (GCP) environment managed and maintained by GovTech, keeping training data isolated. The second is no citizen data: private and personally identifiable information was excluded from training, which used only anonymized datasets derived from scam reports and intelligence feeds. The third is stress-testing: the model was tested with synthetic dialogues and local corpora against diverse scam categories and evolving impersonation tactics. By tying training to a government-managed environment and finishing the judgment on the user's device, the design answers with engineering the question of whether protection requires handing conversations to someone else.

The model sits on top of defenses Singapore already runs. According to the public speech text released by the Cyber Security Agency of Singapore and reviewed by METAL, SATIS, the Scam Analytics and Tactical Intervention System built jointly by GovTech and the Singapore Police Force, scans more than 600,000 websites a day in a matter of seconds each and disrupts around 20,000 malicious websites a month. Its scope has expanded to suspicious mobile numbers and online platforms. Information passed on by the police helped Meta remove or restrict nearly 3.8 million scam-linked entities, pages and accounts on Facebook and Instagram. The ASEAN Guide on Anti-Scam Policies and Best Practices, initiated by Singapore, was endorsed by member states this year. Pointing to a scam chain that moves a victim from social media to a messenger, a phone call and finally a transfer through a financial institution, Teo said "no single defender sees the whole chain."

The partnership's roots go back to the spring. According to reports, the model is a product of the National AI Partnership with the Singapore Government that Google DeepMind announced on May 20, and DeepMind opened a research lab in Singapore the year before. Under the same partnership, work is also under way on exploring AI co-clinician collaboration with public health clusters, a Gemma-powered running assistant for blind and low-vision athletes, and research on multilingual safety benchmarks. On X, DeepMind said that "by fine-tuning it on local scam tactics, it identifies threats in real time without compromising personal privacy."

What remains is the path to deployment. According to reports, no release date has been disclosed, and it is not known whether the model will become a default feature of the Android operating system. DeepMind said messaging platforms and operating system vendors can integrate these lightweight models directly to protect users on the device. Stopping crimes that unfold in conversations servers cannot see is shifting to judgments made inside the handset, and how many phones carry that judgment will depend not on the government or the model provider but on the choices of the companies that control messengers and operating systems.